yingqianwang / oacc-net

[CVPR 2022] Occlusion-Aware Cost Constructor for Light Field Depth Estimation

🌱News🌱

• 2022-03-02- Our paper "Occlusion-Aware Cost Constructor for Light Field Depth Estimation" is accepted by CVPR 2022. Code is available at OACC-Net.
• 2022-02-16- Our paper "Disentangling Light Fields for Super-Resolution and Disparity Estimation" is accepted by IEEE TPAMI.
• 2021-12-25- The source codes of our DistgSSR, DistgASR, and DistgDisp have been released.
• 2021-12-10- An online review of light field image super-resolution is available at LF-Image-SR, and an open source toolbox is available at ZhengyuLiang24/BasicLFSR.
  
  

oacc-net's Issues

在自己拍摄的数据上进行测试

请问为什么我在使用自己拍摄的图像上进行测试时无法获得非常好的效果，在数据获取上有什么注意事项吗？

Is there a 7*7 Angle resolution pre-training model?

Dude, you're doing a great job. Is there a 7*7 Angle resolution pre-training model?

请问得到.pfm深度图,每个像素的距离是多少?

很棒的项目,请问得到.pfm深度图,每个像素的距离是多少?有人可以回复我一下吗

about test time

Hello, I run the test_inference_time.py to reproduce the test time, but I found the warp operation in Generate_mask() costs about 10s, it makes me confused. I wonder if you ever encountered such a problem, thank you.

can not reimplement the reported result, can you help me?

i try to train the model on a RTX 3090 with following hyper-parameters around 1000 epochs, it turns out that the MSE metric cannot go up higher, is this a normal phenomenon?

Multithreading gets stuck during training

Hi, buddy. Your work is excellent. When I try to duplicate your code, I find the code gets stuck. I carefully check the code, finally I find it may be caused by while (sum_diff < 0.01 or glass_region == True): in the function __getitem__ . When I annotate the line code, the training process is normal. I don't know if my behavior is correct. Please give me some suggestions. Thx.